TensorBuffer class

A multi-dimensional array view over typed data with shape and stride metadata.

TensorBuffer provides a NumPy-like interface for tensor operations. It uses a view/storage separation pattern where TensorStorage holds the physical data and TensorBuffer defines how to interpret it through shape and stride information.

This design enables O(1) operations like transpose by manipulating strides rather than copying data.

// Create a 3x4 tensor filled with zeros
final tensor = TensorBuffer.zeros([3, 4], dtype: DType.float32);

// Transpose without copying data
final transposed = tensor.transpose([1, 0]); // Now 4x3

// Reshape (requires contiguous memory)
final reshaped = tensor.reshape([2, 6]);

Available extensions

Constructors

TensorBuffer({required TensorStorage storage, required List<int> shape, List<int>? strides, int storageOffset = 0, MemoryFormat memoryFormat = MemoryFormat.contiguous}): Creates a new tensor buffer with the given storage, shape, and optional strides.

Properties

data → TypedData: The underlying typed data for direct access.
no setter
dataAsFloat32List → Float32List: The underlying data as a Float32List.
no setter
dtype → DType: The data type of elements in this tensor.
no setter
hashCode → int: The hash code for this object.
no setterinherited
isContiguous → bool: Whether this tensor's data is stored contiguously in memory.
no setter
isViewable → bool: Available on TensorBuffer, provided by the TensorViewExtension extension
Checks if this tensor can be represented as a contiguous view of another tensor's storage.
no setter
memoryFormat → MemoryFormat: The memory layout format of this tensor.
final
numel → int: The total number of elements in this tensor.
no setter
rank → int: The number of dimensions in this tensor.
no setter
runtimeType → Type: A representation of the runtime type of the object.
no setterinherited
shape → List<int>: The dimensions of this tensor.
final
sizeInBytes → int: The total size of this tensor's data in bytes.
no setter
storage → TensorStorage: The underlying storage containing the physical data.
final
storageOffset → int: The offset into storage where this tensor's data begins.
final
strides → List<int>: The number of elements to skip in storage for each dimension.
final

Methods

clone() → TensorBuffer: Creates a deep copy of this tensor with its own storage.
contiguous() → TensorBuffer: Returns a contiguous copy of this tensor if not already contiguous.
flatten() → TensorBuffer: Available on TensorBuffer, provided by the TensorViewExtension extension
Creates a flat view of this tensor.
max() → double: Available on TensorBuffer, provided by the TensorBufferReduce extension
Returns the maximum value among all elements in this tensor.
maxAxes(List<int> axes, {bool keepDims = false}) → TensorBuffer: Available on TensorBuffer, provided by the TensorBufferReduce extension
Returns a tensor with the maximum value along multiple axes.
maxAxis(int axis, {bool keepDims = false}) → TensorBuffer: Available on TensorBuffer, provided by the TensorBufferReduce extension
Returns a tensor with the maximum value along the specified axis.
mean() → double: Available on TensorBuffer, provided by the TensorBufferReduce extension
Returns the arithmetic mean of all elements in this tensor.
meanAxes(List<int> axes, {bool keepDims = false}) → TensorBuffer: Available on TensorBuffer, provided by the TensorBufferReduce extension
Returns a tensor with the mean of elements along multiple axes.
meanAxis(int axis, {bool keepDims = false}) → TensorBuffer: Available on TensorBuffer, provided by the TensorBufferReduce extension
Returns a tensor with the mean of elements along the specified axis.
min() → double: Available on TensorBuffer, provided by the TensorBufferReduce extension
Returns the minimum value among all elements in this tensor.
minAxes(List<int> axes, {bool keepDims = false}) → TensorBuffer: Available on TensorBuffer, provided by the TensorBufferReduce extension
Returns a tensor with the minimum value along multiple axes.
minAxis(int axis, {bool keepDims = false}) → TensorBuffer: Available on TensorBuffer, provided by the TensorBufferReduce extension
Returns a tensor with the minimum value along the specified axis.
narrow(int dim, int start, int length) → TensorBuffer: Available on TensorBuffer, provided by the TensorViewExtension extension
Narrows the tensor along a dimension.
noSuchMethod(Invocation invocation) → dynamic: Invoked when a nonexistent method or property is accessed.
inherited
reshape(List<int> newShape) → TensorBuffer: Returns a view of this tensor with a new newShape.
select(int dim, int index) → TensorBuffer: Available on TensorBuffer, provided by the TensorViewExtension extension
Selects a single index along a dimension, returning a view with reduced rank.
sliceFirst(int start, int end) → TensorBuffer: Available on TensorBuffer, provided by the TensorViewExtension extension
Creates a view of a contiguous slice along the first dimension.
squeeze([int? dim]) → TensorBuffer: Returns a view with all size-1 dimensions removed.
sum() → double: Available on TensorBuffer, provided by the TensorBufferReduce extension
Returns the sum of all elements in this tensor.
sumAxes(List<int> axes, {bool keepDims = false}) → TensorBuffer: Available on TensorBuffer, provided by the TensorBufferReduce extension
Returns a tensor with the sum of elements along multiple axes.
sumAxis(int axis, {bool keepDims = false}) → TensorBuffer: Available on TensorBuffer, provided by the TensorBufferReduce extension
Returns a tensor with the sum of elements along the specified axis.
toChannelsFirst() → TensorBuffer: Available on TensorBuffer, provided by the TensorViewExtension extension
For a 4D NHWC tensor, converts to NCHW format without copying.
toChannelsLast() → TensorBuffer: Available on TensorBuffer, provided by the TensorViewExtension extension
Returns a view with channels moved to a different position.
toList() → List<double>: Available on TensorBuffer, provided by the TensorBufferReduce extension
Returns all elements as a List<double>.
topk(int k, {int axis = -1, bool largest = true, bool sorted = true}) → TopKResult: Available on TensorBuffer, provided by the TopKExtension extension
Returns the k largest or smallest values and their indices along axis.
toString() → String: A string representation of this object.
override
transpose(List<int> axes) → TensorBuffer: Returns a view of this tensor with dimensions permuted according to axes.
unbind(int dim) → List<TensorBuffer>: Available on TensorBuffer, provided by the TensorViewExtension extension
Splits the tensor along a dimension into multiple views.
unsqueeze(int dim) → TensorBuffer: Returns a view with a size-1 dimension inserted at position dim.

Operators

operator ==(Object other) → bool: The equality operator.
inherited
operator [](List<int> indices) → double: Returns the element at the given multi-dimensional indices.

Static Methods

arange({required double start, required double end, double step = 1.0, DType dtype = DType.float32}) → TensorBuffer: Creates a 1D tensor with values in a range with a given step.
computeStrides(List<int> shape, MemoryFormat format) → List<int>: Computes strides for a tensor with the given shape and format.
eye(int n, {int? m, DType dtype = DType.float32}) → TensorBuffer: Creates a 2D identity matrix.
fromFloat32List(Float32List data, List<int> shape) → TensorBuffer: Creates a tensor from an existing Float32List with the given shape.
fromFloat64List(Float64List data, List<int> shape) → TensorBuffer: Creates a tensor from an existing Float64List with the given shape.
fromUint8List(Uint8List data, List<int> shape) → TensorBuffer: Creates a tensor from an existing Uint8List with the given shape.
full(List<int> shape, {required double fillValue, DType dtype = DType.float32}) → TensorBuffer: Creates a tensor filled with a specific value.
linspace(double start, double end, {required int steps, DType dtype = DType.float32}) → TensorBuffer: Creates a 1D tensor with evenly spaced values.
ones(List<int> shape, {DType dtype = DType.float32}) → TensorBuffer: Creates a tensor filled with ones.
randn(List<int> shape, {DType dtype = DType.float32, int? seed}) → TensorBuffer: Creates a tensor with random values from a standard normal distribution N(0, 1).
random(List<int> shape, {DType dtype = DType.float32, int? seed}) → TensorBuffer: Creates a tensor with random values uniformly distributed in [0, 1).
uninitialized(List<int> shape, {DType dtype = DType.float32, MemoryFormat memoryFormat = MemoryFormat.contiguous}) → TensorBuffer: Creates a tensor buffer without initializing values.
zeros(List<int> shape, {DType dtype = DType.float32, MemoryFormat memoryFormat = MemoryFormat.contiguous}) → TensorBuffer: Creates a tensor filled with zeros.